U of T team builds culturally aware AI assistant for Bangladesh mental health workers

Why standard language models fall short

The project starts from a language problem. In Bangladesh, a community mental health worker may sit with a client who cannot sleep or stop worrying, and who describes the feeling with the Bangla word manoshik chap, which translates as "tension." According to the Schwartz Reisman Institute, the word carries economic, social and spiritual meanings, and almost none of them fit the categories that today's most powerful AI systems were trained to recognize.

A large language model that flattens these meanings into "anxiety" or "depression" loses nuance and can misinterpret a situation entirely, the article says. "In mental health, the stakes are higher because what is being translated is a person's own account of their suffering," says Ahmed, an associate professor in the Department of Computer Science and a Schwartz Reisman Institute faculty affiliate.

The need is large. The source states that Bangladesh has roughly one therapist for every 85,000 people. Ahmed says mental health is often neglected across the Global South, and that existing tools "don't always work outside of the Western bubble where they were created."

A tool for para-counselors, not for patients

The Wellcome Trust grant of $1.8 million will fund an assistant for para-counselors, a type of frontline community health worker who is often the only source of mental health support in low-resource settings across Bangladesh. Wellcome is a global charitable foundation, and its Mental Health program supports research on anxiety, depression and psychosis.

Counselors will use the assistant to prepare for sessions, make sense of what they heard from patients and think through next steps. The system never speaks directly to patients. The counselor reviews, edits or discards every suggestion, with the goal of strengthening their judgment rather than replacing it.

Ahmed contrasts this with apps where a person in distress talks directly to an unsupervised model. "We would never accept that from a human counselor. We should not accept it from a machine either," he says. In his words, "The counselor keeps the relationship, the accountability and the final word."

Partners, design and timeline

During the two-year project, Ahmed's Third Space research group at U of T will work with two Bangladeshi partners: the non-profit SAJIDA Foundation and the for-profit social enterprise mPower Social Enterprises Limited. A network of people with lived experience of anxiety, depression and psychosis will also take part.

The project treats ethics and data sovereignty as central design questions. People with lived experience act as co-designers, reviewers and safety monitors, and are fairly compensated for their expertise. The underlying models and methods will be released openly so other communities can adapt them.

The team plans to do the following:

• assemble an ethically sourced body of data
• build and test the AI
• co-design the tool with the counselors who will use it
• run a six-month field pilot across five regions of Bangladesh

Arid Hasan, a PhD student supervised by Ahmed, has led the research team for seven months and has worked on Bangla natural language processing since 2018. Graduate and undergraduate students on the project will gain hands-on training in building culturally grounded machine learning and language systems.

What it could mean beyond Bangladesh

The source describes the system as poised to become one of the first AI mental health interventions designed specifically for communities in the Global South. If the pilot succeeds, the results could extend to other languages and places such as West Bengal, Nepal and East Africa.

Ahmed frames the stakes broadly. "If we get this right, the contribution is not only a tool," he says. "It's a demonstration that the Global South can be where responsible AI is built, not just where it is deployed after the fact."